Crimeless

Mission

Heatmap of Crime in Boston

Who We Are

We are a group of 4 seniors at Harvard College with a strong interest in using data science to solve real-world problems.

What We Hope to Achieve

To date, “Big Data” and Data Analysis have helped corporations become more cost-effective and politicians reach voters in smarter ways. Unfortunately, despite the ubiquity of publicly available data, there haven’t been large-scale or successful attempts at helping Law Enforcement police the streets more effectively, and those attempts that have been made are often on expensive platforms that are not readily accessible to the general public.

What we seek to do in this project is build out data analytics models for Law Enforcement that can demonstrate the power of data analysis effectively enough so as to encourage more work on this topic and to keep our streets safe.

Development

The Making of Crimeless

The first part of our project is our Static Analysis. This part of our project consists of an exploratory study of crime in Boston and an analysis of which static factors (i.e. factors that do not change on an hourly or daily basis) are most likely to lead to a more crime-prone area.

The second part of our project is our Dynamic Analysis. This part of the project consists of predicting when and where a crime will occur by using dynamic factors (things that change frequently over time such as weather, time of day, public holidays or weekends etc) to determine if a crime will occur or not in a given location at a given time. We then use the accuracy of our predictions and a probability of success at stopping a predicted crime to estimate an estimated positive economic impact on Boston.

Static Data Processing

We begin development with our Static Data Processing, which we start by pulling Boston’s Public Datasets on Crime Incident Reports from July 2012 to August 2015 and proceed to clean, post-process and quantify the data yielding a total of 250,000 reported incidences. We then also pull various static variables that we believe may be correlated to the number and severity of crimes that occur in each of these locations, including their average property value, amount of liquor stores, and quantity of streetlights.

Static Data Modeling

Once we have our data cleaned and visualized we can move on to our Static Data Modeling. In this part we try to predict a danger index, which we calculate as being the sum of the number of crimes and their severities in an area, of a particular location based on the static variables described in previous section. We try several different classifiers here and use the best one as our predicted danger rating for every location in our dataset.

Dynamic Data Processing

This part of our project starts with our Dynamic Data Processing. This involves pulling dynamic data factors that we think might help us determine if a crime will occur at a given location and time such as: temperature, visibility, humidity, precipitation, time of day, day of the week, unemployment rate, and whether it’s a public holiday or weekend. We pull this data from various APIs, public and paid data sources and proceed onwards.

Dynamic Data Modeling

Our project then proceeds to our Dynamic Model. In this part, we train several different models such as SVMs, Linear and Logistic Regression, Naive Bayes and Random Forest to try and predict if a crime occurs at a location and time. We optimize our learners and then use our predictions and the economic costs of crimes in our datasets to get a measure of the positive economic impact our predictions could have potentially had on Boston if they were acted on.

Analysis

Our project attempts to predict whether or not a crime will occur in a given location at a given time based on a set of dynamic factors in conjunction with static factors that we processed earlier. Ultimately, this type of data could be incredibly helpful for police in order to dynamically allocate their patrols to the most threat-prone areas based on the changing variables of the day. Based our analysis, we were able to find out that the Random Forest Classifier optimized over our training dat does a fantastic job at saving the Boston Police Department money. There are a number of considerations that one should look into before "blindly" accepting our results -- it may be a an oversimplification to say that the cost of additional police is just the additional hourly wage paid to the officers, and it may also be an optimistic oversimplification to say that if cops know when/where a crime is going to occur, they can prevent it with the probabilities discussed above. However, none the less, we feel that the vast majority of our assumptions were grounded in intensive research and are fair to be considered over a vast range of potential scenarios. At the end of the day, the number we get of saving the Boston Police Department a potential of 67 million dollars over 3 and a half years is something we believe at least indicates that this is a direction the department could benefit from exploring further and hopefully invites future work into the space.

Crimeless

Our Team

The Creators of Crimeless

Dev Rishi

Carl Fernandes

James Ruben

Wilder Wohns

Mission

Who We Are

What We Hope to Achieve

Development

The Making of Crimeless

Static Data Processing

Static Data Modeling

Dynamic Data Processing

Dynamic Data Modeling

Analysis

Future Work

Utilization of More Factors

Improved Model Performance

Testing in Other Cities

Contact Us

Leave us a message